Google Launches Tool Retrieval in Gemini API: On-Demand Tool Schemas to Cut Token Costs
⚡ Breaking News
Google AI Blog
November 20, 20264 min read0

Google Launches Tool Retrieval in Gemini API: On-Demand Tool Schemas to Cut Token Costs

By — Senior AI Technical Editor

Back to News
❝

Google has launched Tool Retrieval in the Gemini API (public preview) within the Interactions API, enabling on-demand loading of tool parameter schemas instead of sending the full tool library with every request. The feature works with Gemini 3.6 and later, including all Gemini 4 models, at no additional charge—tokens are billed at standard model rates. It reduces input token consumption and latency while improving tool selection accuracy as AI agents scale.

Executive Overview

Google has officially announced the launch of Tool Retrieval in the Gemini API, now available in public preview within the Interactions API. The feature allows AI agents to load tool parameter schemas on demand rather than sending the entire tool library with every request. It works with Gemini 3.6 and later models, including all Gemini 4 models, at no additional charge—token usage is billed at standard model rates. By reducing input token consumption and latency while improving tool selection accuracy, Tool Retrieval addresses a core scaling challenge for agentic AI systems.

📊 Official Technical Specifications & Data Sheet

Technical DimensionConfirmed Official Data
💰 Pricing & Usage CostNo separate fees for Tool Retrieval; input and output tokens are billed at standard model rates. Enabling tool_search adds a small fixed block of input tokens, and each loaded schema is billed as input for the remainder of the conversation.
🌐 Platforms & Immediate AvailabilityAvailable now in public preview within the Interactions API via Google's Gemini API; supports remote MCP servers.
⚡ Performance & Speed MetricsReduces input tokens and latency by loading schemas only on demand; improves tool selection accuracy as the candidate list grows; compatible with Prefix Caching.
🛡️ Security & RobustnessSupports user-permission filtering via client-side retrieval, with a detailed log in interaction.steps to track which tools the model considered.
🧠 Context WindowNo new context window is announced; the feature reduces input token consumption within the context window of the model used (Gemini 3.6+ / Gemini 4).
🌍 Arabic Language & Regional SupportThe feature is a developer-facing capability via Gemini API and is not language-specific; Arabic processing support follows the core capabilities of Gemini 3.6 and Gemini 4 models.

Deep-Dive Features & Architecture

Tool Retrieval solves a fundamental problem in building AI agents: when an agent connects to dozens or hundreds of tools, sending every full parameter schema with every request inflates input token counts, increases latency, and degrades tool selection accuracy as the candidate list grows. The new solution keeps the full tool library on standby and exposes only each tool's name and description to the model, while parameter schemas remain outside the context until the model requests them by name. When needed, the model searches for the tool, the API loads its full schema, and the model issues a function_call—all within the same request.

Activation is simple: add the tool_search tool to the request and set defer_loading: true on the tools to be deferred. Developers do not need to build or host an index, because Google performs the search server-side, meaning code changes are limited to the new tool and the deferral flag. Frequently used tools can also be kept preloaded while the long tail is deferred. The feature also supports remote MCP servers: setting defer_loading: true on an mcp_server tool loads every tool the server exposes on demand. Developers who manage a custom tool registry, embedding index, or per-user permission filtering can declare tool_search with "execution": "client", adding one round trip per search.

Benchmark & Competitive Performance

Google does not publish specific benchmark figures for savings percentages in this announcement, but it clearly defines the economic equation: savings are realized when the schemas that are never loaded outnumber those that are actually loaded—the typical situation for large libraries. In practice, Tool Retrieval outperforms the traditional approach of sending all schemas with every request, as well as client-side retrieval solutions that add a round trip per search. Its compatibility with Prefix Caching means the cached prefix remains stable from turn to turn, further improving cost efficiency in long multi-turn conversations.

Industry Impact & Enterprise Adoption

For developers building production-grade AI agents, Tool Retrieval removes a major scaling bottleneck: the trade-off between tool library size and per-request cost. Enterprises deploying agents across large internal tool catalogs—CRM, ticketing, databases, and remote MCP servers—can now expose hundreds of tools without inflating input tokens or latency. The client-side execution option also enables organizations to enforce per-user permission filtering while maintaining a detailed audit trail in interaction.steps, which is critical for regulated industries. Because the feature is available at no additional charge and requires minimal code changes, adoption barriers are low, and it is expected to accelerate the shift toward large-scale, multi-tool agent architectures on Gemini 3.6 and Gemini 4.

Conclusion

Tool Retrieval in the Gemini API is a targeted, high-impact optimization for the agentic era. By loading tool schemas on demand, Google reduces input token consumption and latency while improving tool selection accuracy—without introducing new fees or complex infrastructure requirements. Available now in public preview within the Interactions API and compatible with remote MCP servers and Prefix Caching, the feature gives developers a practical path to scale AI agents across large tool libraries. As agent ecosystems grow, on-demand schema retrieval is likely to become a standard component of production AI stacks.

Media Source: Google AI Blog | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis

Original Source:Google AI BlogThis news was formulated based on coverage from Google AI Blog

Frequently Asked Questions

What is the Tool Retrieval feature in Gemini API?

It is a new feature in public preview within the Interactions API that allows an agent to keep a full tool library on standby and load parameter schemas only when needed, instead of sending all schemas with every request. It is enabled by adding the tool_search tool and setting defer_loading: true.

What is the cost of using Tool Retrieval in Gemini API?

There are no separate fees for Tool Retrieval. Input and output tokens are billed at the standard rates of the model used. Enabling tool_search adds a small fixed block of input tokens, and each schema loaded by the model is billed as input for the remainder of the conversation.

Which models support Tool Retrieval?

The feature works with Gemini 3.6 and later versions, including all Gemini 4 models, within the Interactions API.

Is Tool Retrieval available via remote MCP servers?

Yes, the feature works with remote MCP servers in the Gemini API. Setting defer_loading: true on an mcp_server tool loads each tool exposed by the server on demand without listing them one by one.

How does Tool Retrieval interact with Prefix Caching?

The feature is compatible with prefix caching: deferred schemas remain outside the start of the request, and schemas loaded by the model are added later in the conversation, keeping the cached prefix stable from turn to turn.

AI Tools Oasis

AI Tools Oasis Team

Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.